fix(mcp): isolate stdio wire protocol, add cache-first embedder fast path and background warmup

[Coding-Dev-Tools]

Summary

Fixes low-level stream write failures and child process disconnection errors (such as \Error: write EOF\ / \UV_EOF\ at \WriteWrap.onWriteComplete) encountered by Node.js MCP harnesses (Command Code, Cursor, Claude Code, Cline, etc.) when calling Engraphis tools.

Root Causes & Solutions

1. Stdio Wire Protocol Contamination:

   • In MCP stdio mode, standard output is the JSON-RPC wire. Third-party dependency logs, warnings, or progress bars (PyTorch, transformers, tqdm, pydantic) writing to stdout corrupted the JSON-RPC wire, causing client parsers to fail and close the stdin pipe.
   • Fix: Redirected \sys.stdout\ to \sys.stderr\ in \engraphis.mcp_server\ during stdio mode while preserving the raw binary stream specifically for JSON-RPC transport.

2. First-Call Cold Start & Network Hangs:

   • Lazy initialization meant the first tool call (such as
     ecall) paid the full initialization and network check cost inside the agent's turn. Hugging Face checks could hang on restricted networks.
   • Fix:
     • Added cache-first \local_files_only=True\ fast path in \get_embedder\ so cached models load in ~0.3s without network calls.
     • Made \service()\ singleton initialization thread-safe (_service_lock = threading.Lock()).
     • Added non-blocking background daemon warmup (_start_background_warmup()) in \mcp_server.py\ so SQLite and the embedder are already warm before the first tool call.

3. Pre-flight Health Verification & Prefetching:

   • Added embedder health check to \engraphis-init --check\ (validating PyTorch C++ DLL loading on Windows).
   • Added \engraphis-init --prefetch\ to download and cache model weights during setup.

Verification

• Unit tests added in \ ests/test_mcp_server.py, \ ests/test_backends_factories.py, and \ ests/test_init.py.
• 
  uff check .\ passed with 0 errors.
• \pyright\ passed with 0 errors.
• \check_commercial_manifest.py\ and \externalize_dashboard_assets.py\ passed.
• Full offline test suite (\pytest tests/) passed 100%.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.